Add support for float type raster datasets

[tritolol]

I'm trying to load a custom DSM raster dataset of type float32.

I'm implementing a new class based on RasterDataset like so:

from torchgeo.datasets import RasterDataset

class DsmData(RasterDataset):
    filename_glob = "*.tif"

When I call __getitem__() on a DsmData object, I get an array containing int32 data instead.
I traced this down to the following line in the RasterDataset definition

torchgeo/torchgeo/datasets/geo.py

Line 464 in 3f7e525

dest = dest.astype(np.int32)

All raster datasets are forced into the int32 type which should not happen.

Having RasterDatasets with different types will probably cause problems when applying union or intersection operations.
But since a custom collate_fn can be defined, the user is able to provide a solution for this.

I achieved the desired behavior simply by removing the mentioned line.

[calebrob6]

Oof, good catch. We should definitely change this. Do you want to open a PR to get this started?

[adamjstewart]

Having RasterDatasets with different types will probably cause problems when applying union or intersection operations.

Nope, this won't be a problem. All of our collation functions use torch operators that can handle various dtypes:

$ python
>>> import torch
>>> a = torch.zeros(3, 3, dtype=torch.int32)
>>> b = torch.zeros(3, 3, dtype=torch.float32)
>>> torch.stack((a, b))
tensor([[[0., 0., 0.],
         [0., 0., 0.],
         [0., 0., 0.]],

        [[0., 0., 0.],
         [0., 0., 0.],
         [0., 0., 0.]]])
>>> torch.maximum(a, b)
tensor([[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]])
>>> torch.cat((a, b))
tensor([[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]])

[tritolol]

Created PR #384